RedPen
An offline contract review tool for freelancers. It uses a two-layer architecture combining a rules engine with a small model, runs entirely locally, and is sold for a one-time fee. It can identify risky clauses offline and generate customer notification emails.
Timeline (last 90 days)
RedPen is an offline contract review tool for freelancers, whose core design philosophy is that the model never judges what clauses are risky, 65 open-source rules are responsible for identifying dangerous clauses, and a locally running Gemma 3 1B model is only responsible for reorganizing the rule findings into language that can be sent to clients.
The entire RedPen tool runs on the user's local laptop, uploads no data, charges no per-use fee, and is priced at a one-time $19.
Even with the model layer completely turned off, RedPen's review report remains fully usable.
RedPen's rule layer is written in pure Python and is divided into three rule libraries—freelance agreements, offer letters/internship agreements, and NDAs—with 65 open-source rules in total, and each rule carries its own explanation, a ready-to-paste counter-clause, and what to say in the email.
RedPen uses PyMuPDF for text extraction; the segmentation module restores clause numbering and preserves each clause's page number and character span, and every finding quotes verbatim the source text that triggered it.
RedPen runs on an 8GB Apple Silicon MacBook, requiring no GPU and no Docker; Gemma 3 1B is called via llama.cpp after Q4 quantization, with weights of 769MB, a load time of 2.3 seconds, and a resident memory of 0.83GB; rule processing takes 12-28 milliseconds, a full review (including the model layer) takes 11-17 seconds, and results are rendered via SSE streaming.
In tests on six real documents, RedPen exposed six bugs: silently dropping page 1 of an offer letter, incorrectly reporting a no-deposit clause for an intern, missing an impossible date, missing a disguised perpetual clause, flagging normal drafting as risky, and treating blank fields as hostile.
RedPen's guardrails include: when citing a clause that does not exist, resolving it against the real clauses or otherwise discarding it; checking coverage to ensure wording requested by the client is not omitted; stripping and replacing the email signature with [your name]; reading labels from the actually running model rather than hardcoding them; and having a deterministic fallback for each role to guarantee the report is complete.
RedPen enforces network isolation at the code level: the HTTP client throws an exception when the base URL is not loopback, with test coverage; uploaded files are parsed from bytes and never written to disk; scripts/offline_proof.sh samples process sockets and verifies they are all loopback.
One structural advantage of RedPen choosing to run locally is zero marginal cost: hosting a server that can hold the weights costs about $25/month, runs at only 2-5 tokens/second on a shared CPU—slower than the user's own laptop—and every contract would end up on a server outside their control; with local execution the 500th user costs zero, which is the basis for the $19 one-time pricing rather than a subscription.
4 more timeline events
All Facts (14)
RedPen is an offline contract review tool for freelancers, whose core design philosophy is that the model never judges what clauses are risky, 65 open-source rules are responsible for identifying dangerous clauses, and a locally running Gemma 3 1B model is only responsible for reorganizing the rule findings into language that can be sent to clients.
50%UnverifiedThe entire RedPen tool runs on the user's local laptop, uploads no data, charges no per-use fee, and is priced at a one-time $19.
50%UnverifiedEven with the model layer completely turned off, RedPen's review report remains fully usable.
50%UnverifiedRedPen's rule layer is written in pure Python and is divided into three rule libraries—freelance agreements, offer letters/internship agreements, and NDAs—with 65 open-source rules in total, and each rule carries its own explanation, a ready-to-paste counter-clause, and what to say in the email.
50%UnverifiedRedPen uses PyMuPDF for text extraction; the segmentation module restores clause numbering and preserves each clause's page number and character span, and every finding quotes verbatim the source text that triggered it.
50%UnverifiedRedPen runs on an 8GB Apple Silicon MacBook, requiring no GPU and no Docker; Gemma 3 1B is called via llama.cpp after Q4 quantization, with weights of 769MB, a load time of 2.3 seconds, and a resident memory of 0.83GB; rule processing takes 12-28 milliseconds, a full review (including the model layer) takes 11-17 seconds, and results are rendered via SSE streaming.
50%UnverifiedOne structural advantage of RedPen choosing to run locally is zero marginal cost: hosting a server that can hold the weights costs about $25/month, runs at only 2-5 tokens/second on a shared CPU—slower than the user's own laptop—and every contract would end up on a server outside their control; with local execution the 500th user costs zero, which is the basis for the $19 one-time pricing rather than a subscription.
50%UnverifiedIt has been verified that the same contract produces completely identical findings under Gemma 3 1B and Qwen 2.5 1.5B—same rules, same pages, same evidence—with only the summary and email wording changing.
50%UnverifiedRedPen's hardware floor is 8GB of memory, at which 1-3B Q4-quantized models can run and 7B cannot; scanned PDFs are not supported, and when there is no selectable text the file is not sent to an OCR service but simply stated as unsupported.
50%UnverifiedRedPen's guardrails include: when citing a clause that does not exist, resolving it against the real clauses or otherwise discarding it; checking coverage to ensure wording requested by the client is not omitted; stripping and replacing the email signature with [your name]; reading labels from the actually running model rather than hardcoding them; and having a deterministic fallback for each role to guarantee the report is complete.
50%UnverifiedRedPen enforces network isolation at the code level: the HTTP client throws an exception when the base URL is not loopback, with test coverage; uploaded files are parsed from bytes and never written to disk; scripts/offline_proof.sh samples process sockets and verifies they are all loopback.
50%UnverifiedIn tests on six real documents, RedPen exposed six bugs: silently dropping page 1 of an offer letter, incorrectly reporting a no-deposit clause for an intern, missing an impossible date, missing a disguised perpetual clause, flagging normal drafting as risky, and treating blank fields as hostile.
50%UnverifiedThe Actizant founding team, after reviewing the codebase, gave a three-line roadmap: profession-specific risk rules, jurisdiction-aware guidance, and transparently declaring the limitations of automated analysis.
50%UnverifiedFreelance developer Shreshtha Das tested RedPen and gave positive feedback, particularly emphasizing the advantage of offline operation when handling confidential client agreements and NDAs; the developer considers this a single data point, not validation, and the false positives he most wanted were not reported by the other party.
50%